Bürokratt: Estonia's single network of AI assistants replacing dozens of separate agency chatbots

The problem

Estonia had built a wide range of e-government services, but the ways people accessed them were fragmented. Different government agencies had each developed their own chatbots, leaving citizens and businesses to work out which service sat behind which website and which assistant to ask. The more services moved online, the harder it became to navigate the range of e-government services.

Those earlier assistants were also demanding to run. Each had to be taught a fixed set of questions and answers, so it could handle familiar queries but required constant hands-on updates by the institution behind it, making it resource-intensive to keep up to date.

The underlying need was for a single, consistent way to reach end users, one that could understand an ordinary question in plain Estonian, answer it quickly, and direct the person to the right service, without each agency having to build and maintain its own chatbot.

The solution

Bürokratt is the Estonian state's own AI-based virtual assistant, developed and managed by the Information System Authority (RIA). It takes the form of a connected network of chatbots on public agencies' websites, allowing people to get information and access public services by asking a virtual assistant. It works in plain Estonian, is available at any hour, and hands the user over to a human customer service agent when it cannot help. The longer-term aim is to let people deal with the state through ordinary conversation, getting answers, completing tasks, and using services straight from a chat window on whatever device they have, so they no longer need to find their way around separate portals or fill in forms.

From a single trained chatbot to a network of AI agents

Bürokratt began as a single interoperable chatbot, built on natural language processing and machine learning. It was trained using third-party frameworks, hosted in the national State Cloud, and uses the state authentication service to keep data secure. It is now shifting to large language models, which interpret what a question means and pull answers from sources each institution defines, with no separate manual training. The roadmap sets out a base layer first: a working LLM with retrieval-augmented generation that draws answers from specific databases, a shared knowledge module, and a central classifier that lets the separate Bürokratt assistants securely pass queries between them. After that, the plan is for each institution or policy area to run its own AI agent within a cooperating network, including an LLM adapted to Estonian and agents capable of carrying out tasks on a person's behalf, such as lodging a query, filing an application, or making a booking. Further out, the government is working towards cross-border interoperability, so that assistants in different countries can connect, for instance, linking Bürokratt with Finland's AuroraAI so that an Estonian abroad could arrange services through it.

Built to be cheap to adopt and built on the digital state

The software is free and open source. Hosting runs in the State Cloud at about €150 per institution per month, with additional charges when large language models are used. An institution that wants to join mainly has to gather, and where needed, tidy up, the source data its assistant will rely on, and RIA helps each one along from an initial demo through to going live. Because it runs on the shared state cloud and the national authentication service, data is handled securely across all public bodies that participate. Bürokratt also sits within the responsible-AI requirements Estonia attaches to the AI projects it funds, which the government's chief data officer lists as a data protection impact assessment, cybersecurity checks, data-quality checks, use of the national data tracker that records how personal data is accessed, and purpose limitation.

The impact

Bürokratt has moved from development into live use, so the picture so far is mainly one of adoption and expected benefits, with few published outcome metrics yet.

1. It is used across many public institutions

Several public sector organisations already run Bürokratt, among them the Tax and Customs Board, the Health Insurance Fund, the Police and Border Guard Board, Statistics Estonia, the Environment Agency, and others spanning national agencies and local government, such as Rae Parish. By October 2025, the government's chief data officer put the total at 18 organisations, with more expected.

2. The move to large language models is already cutting maintenance

RIA's own electronic identity service, id.ee, was the first to run the LLM version, which answers genuine user questions without manual training and, by RIA's account, provides better answers with less upkeep than the older chatbots.

3. The expected gains are efficiency and access

Customer support staff in participating organisations are expected to handle lower volumes of routine queries, freeing time for other work, while citizens are expected to reach services and information more quickly through a single, more interoperable point of contact.

Lessons learned

• One front door beats many separate bots. Agency-by-agency chatbots left people unsure where to turn. Bürokratt's value is in being a single, interoperable network that gives a consistent way to reach the state across institutions.

• Large language models can lift the maintenance burden. Earlier chatbots needed constant manual question-and-answer training. Moving to LLMs that draw on institution-defined sources reduces upkeep and tends to yield better answers, which is a main reason institutions are switching.

• Make adoption cheap, but be honest about the data work. Free open-source software and modest hosting costs lower the barrier to entry, but institutions still have to collect and clean the data the assistant relies on. Central support from RIA, from an early demo through to going live, helps them over that hurdle.

• Coordination is the hard part. Building a shared AI solution across many organisations requires real expertise, and not every institution is ready to participate. The programme has openly described engaging partners in the network as one of its genuine difficulties.

• Open, modular foundations let others start small and scale. Estonia's chief data officer points to Bürokratt's modular design as what lets an administration begin with something simple, such as notifications, and expand towards proactive, event-driven services as it is ready. By his account, what transfers to other governments is the set of foundational principles, open interfaces, built-in identity and consent, local-language models, and clear guardrails for trust, more than the specific technology.

• Build on the digital-state foundation. Running inside the national cloud and using the state authentication service lets Bürokratt handle data securely across public administrations, so each agency does not have to solve identity and data protection on its own.

Launch year: 2021

This case study was written with assistance from artificial intelligence.